KR101715709B1 - Noise reduction of breathing signals - Google Patents

Abstract

The present invention relates to a system and method for processing respiratory signals. A noise reduction operation is performed on the spectral respiration signal 18 to compute an output spectral signal 38, which noise reduction operation uses a spectral subtraction and a gain function < RTI ID = 0.0 > Dimensional frequency and temporal filtering of the gain function, e. G., Two-dimensional frequency and temporal filtering of the gain function, is performed. For example, the spectral respiration signal is calculated based on the respiration signal.

Description

{NOISE REDUCTION OF BREATHING SIGNALS}

The present invention relates to the field of processing breathing signals, in particular, acoustic breathing signals, and more particularly, to a method of processing breathing signals and a system for processing breathing signals.

Respiratory signals or respiratory signals and respiratory rates are basic vital signs. For example, respiration rate may be obtained from measured respiratory waveforms. For example, respiratory waveform signals are generated through sensor electrodes externally attached to the person to be respirated. The breathing waveform may also be derived from an electrocardiogram (ECG) waveform.

US 6 290 654 B1 discloses a device for detecting the breathing pattern of a breathing patient. An aerial microphone is used to sense breathing or snoring sounds. Audible sound signals from the patient ' s body are converted into electrical signals by the microphone and other organ microphones located in the patient ' s neck, and these electrical signals are supplied to the A / D converter. The output from each A / D converter is coupled to a set of delay line elements, each of which is represented by one sampling period of delay, and an active linear filter that uses an adaptive linear filter consisting of a corresponding set of adjustable coefficients Is supplied to the noise cancellation unit. The output of the adaptive linear filter is subtracted from the sensor output. The resulting output is used to adjust the tap weights in the adaptive linear filter to minimize the mean squared value of the total output. The signal noise from the engine microphone is handled in a manner similar to the signal noise from the atmospheric microphone. The signals are band-pass filtered and supplied to the estimated volume of the airflow waveform generation unit.

Measurements that require the placement of microphones or sensor electrodes attached to a person are realistically visible and uncomfortable to humans. Specifically, this is important when the respiratory signals of the sleeping person are to be processed.

It is therefore desirable to be able to process respiration signals captured by a microphone located a great distance from the person breathing.

For example, when recording respiratory signals by a microphone, the microphone may be placed near the respirator, e.g., 50 cm away from the respirator. However, since the acoustic energy resulting from breathing is generally very weak, the signal-to-noise ratio, i.e. the ratio between the breathing signal and the noise, can be very low, making it difficult to extract the relevant respiratory parameters such as respiration rate from the signal.

It is desirable to be able to perform a new kind of noise reduction operation to facilitate determining respiratory parameters when capturing acoustic respiration signals away from the person sleeping.

It has been found that most microphones with amplifiers built into the microphone exhibit sensor noise that exhibits low-pass characteristics due to the 1 / f-noise of the amplifier. Some microphones, such as micro-electromechanical system (MEMS) sensors, have a signal-to-noise ratio (SNR), or signal-to-noise ratio, of approximately 60 dB. In addition, good microphones may have an SNR range of 70 to 80 dB.

It is desirable to provide a method or system for processing respiratory signals, which method facilitates reducing sensor noise from the microphone.

In order to better address one or more of these problems, in a first aspect of the present invention, a method of processing respiratory signals is provided,

Performing a noise reduction operation on the spectral respiration signal to compute an output spectral signal, the noise reduction operation using spectral subtraction; performing the noise reduction operation; And

Performing two-dimensional frequency and time filtering of the gain function used in the spectral subtraction of the noise reduction operation performing step.

For example, the noise reduction operation includes the step of performing the two-dimensional filtering of the gain function.

For example,

- calculating the spectral respiration signal based on the respiration signal, e.g. based on the respiration signal in the time domain.

For example, the breathing signal is an acoustic breathing signal. For example, a breathing signal is a signal captured by a microphone.

For example, calculating the spectral respiration signal includes a fast Fourier transform (FFT).

The term "noise reduction operation" should be understood to mean an operation suitable for reducing the noise contained in the signal represented by the spectral respiration signal, e.g., sensor noise.

The term "filtering" should be understood to include average, median filtering, and other filtering algorithms.

In a noise reduction operation using spectral subtraction, a gain function is used as described in more detail below.

In general, respiration increases and decreases very slowly over time. In addition, the frequency content of the respiration signal does not represent significant peaks and valleys, for example, in the formants of speech signals. It has been found that the spectral characteristics of breath sounds represent related information up to approximately 2 kHz. Since there is a large area in the time-frequency characteristics of the breathing signal that do not represent very large peaks and valleys, two-dimensional frequency and time filtering of the gain function may preferably be applied. Thereby, the noise reduction can be improved while avoiding unacceptable distortion of the breathing signal.

For example, the two-dimensional frequency and time filtering of the gain function are two-dimensional frequency and time averaging / median filtering, i.e., mean or median filtering. In general, filter algorithms such as mean or median filtering are particularly suitable for eliminating outliers.

In one embodiment, the two-dimensional frequency and time filtering of the gain function is two-dimensional frequency and time intermediate filtering. Intermediate filtering is particularly advantageous in that it allows the removal of outliers by replacing the outliers with other values without smoothing the relevant signal portions. Thus, the quality of the noise reduction operation is improved.

For example, the method further includes providing an output signal based on a transformation of the output spectral signal into the time domain. This may be desirable for calculating breathing parameters. For example, the conversion of the output spectrum signal to the time domain may include an inverse fast Fourier transform (IFFT).

For example, the method includes calculating at least one breathing parameter based on the output spectral signal. Important respiratory parameters are, for example, respiratory rates. For example, the at least one breathing parameter is calculated based on the output signal obtained by converting the output spectral signal into a time domain. For example, the at least one breathing parameter is a sleep quality parameter. Thus, sleep quality may be determined. For example, the sleep quality of a sleeping person may be determined or estimated, and a control signal may be provided based on a sleep quality estimate. For example, the control signal may be a signal that controls or influences external factors of the sleep environment, such as temperature, illumination, and the like. For example, one or more control signals may be selected to improve or enhance the sleep of the person being sleeping. Another possibility is to provide a control signal to awaken a person when the person's sleep quality is estimated to be too low, e.g., below the sleep quality threshold.

In one embodiment, the gain function used in the spectral subtraction is calculated based on the magnitude spectrum signal calculated based on the estimated noise and spectral respiration signals. For example, calculating the gain function involves dividing the estimated noise spectrum into an absolute value spectral signal. However, instead of the absolute value spectral signal, for example, a squared absolute value, that is, a power spectral signal, may also be used. Thus, in this case, the gain function is calculated based on the squared estimated noise and magnitude spectrum signal.

For example,

Calculating a magnitude spectrum signal based on the spectral respiration signal;

Calculating a gain function based on the estimated noise and magnitude spectral signal;

Performing two-dimensional frequency and time filtering of the gain function; And

And calculating an output spectral signal by multiplying the spectral respiration signal by a filtered gain function.

That is, the noise reduction operation includes the aforementioned steps of performing the two-dimensional filtering of the gain function. Thus, an efficient noise reduction operation is achieved.

For example, the noise reduction operation may include estimating noise, and the estimation of the noise includes an average operation based on several successive blocks of the magnitude spectrum signal, and the magnitude spectrum signal is calculated based on the spectral respiration signal . For example, the magnitude spectrum signal may be averaged over several successive blocks. Alternatively, for example, the square of the magnitude spectrum signal, i.e., the power spectrum signal, may be averaged. Therefore, a priori knowledge is used that the noise is stationary. This is particularly desirable in situations where, for example, the sensor noise is very stationary, and the noise is mainly determined by the sensor noise of the microphone. For example, an averaging operation is performed for each frequency bin, i. E., For each element of a set of discrete frequencies for which a spectral signal function is defined, e.g., frequency bins of the FFT in particular. For example, the number of contiguous blocks corresponds to a time period of at least 1 second, preferably at least 3 seconds, in particular at least 10 seconds. Thus, a high quality noise floor estimation may be achieved.

For example, the two-dimensional frequency and time filtering of the gain function may be performed using a filter kernel corresponding to a time span of at least 0.05 seconds, more preferably at least 0.1 seconds, or even more preferably at least 0.25 seconds. ).

For example, the two-dimensional frequency and temporal filtering of the gain function is performed with a filter kernel corresponding to at least three consecutive blocks of the magnitude spectrum signal used in the noise reduction operation, and the block size may be, for example, 512 samples admit. More preferably, the number of blocks is at least five, or even more preferably at least seven. Thus, the noise reduction may be improved.

For example, the two-dimensional frequency and time filtering of the gain function is performed with a filter kernel corresponding to a frequency span of at least 40 Hz, more preferably at least 75 Hz, or even more preferably at least 100 Hz. For example, the filter kernel may correspond to a frequency span of at least 250 Hz. For example, the two-dimensional frequency and time filtering of the gain function is performed with a filter kernel corresponding to at least three frequency bins, more preferably at least five frequency bins, or even more preferably at least seven frequency bins . Therefore, the sensor noise may be efficiently removed.

For example, the method includes sampling the breathing signal at a sampling frequency of at least 4 kHz, preferably at least 8 kHz. The use of this relatively high sampling frequency is desirable in that the spectral characteristics of breath sounds represent relevant information up to 2 kHz. Thus, the noise reduction may be improved.

For example, the method includes down-sampling an output signal that is calculated based on the output spectral signal. Using a higher sampling frequency for the noise reduction operation than needed to compute the breathing parameters allows for improved noise reduction. For example, the sampling frequency may be at least 8 kHz. Also, using a high sampling frequency allows two-dimensional filtering of the gain function at frequency and time to be performed without distorting breath sounds.

In one embodiment, the method further comprises adjusting the filter kernel width and / or the filter kernel length of the filter kernel used in the two-dimensional frequency and time filtering of the gain function, The ratio between the signal energy contained in the frequency domain and the energy of the noise energy, for example, the energy of the musical tone included in the second time-frequency domain of the acquired signal based on the two-dimensional frequency and time filtering of the gain function . These signals are, for example, a filtered gain function, a spectral output function, or an output function that is all dependent on the filtered gain function. For example, the signal may be a filtered gain function obtained by two-dimensional frequency and time filtering of the gain function. For example, the first time-frequency domain is comprised of a first frequency domain and a first time domain, and the second time-frequency domain is comprised of a second frequency domain and a second time domain. For example, the first frequency range is a low frequency range and the second frequency range is a high frequency range. The filter kernel width corresponds to the frequency-span, and the filter kernel length corresponds to the time-span of the filter kernel. Preferably, the first and second time-frequency regions are selected so that the desired signal, i.e., the energy of the breathing sound, contributes substantially to the energy contained in the first time-frequency region. Thus, the energy neighbors but contained outside the first time-frequency domain, and in particular the energy contained in the appropriately selected second time-frequency domain, may be due to noise such as a tone.

The terms "signal energy" and "noise energy" refer to the integral squared of the absolute value of the signal amplitude and may be computed in the time domain or frequency domain. That is, the term "noise energy" should be understood as meaning signal energy that is assumed to be attributable to noise.

For example, the adjusting includes optimizing a ratio between signal energy included in a first time-frequency domain of the function and noise energy included in a second time-frequency domain.

For example, the first time range may be selected to correspond to or include the (locally) maximum signal energy of the signal.

For example, the first and second time ranges may not overlap. For example, the first and second time ranges may be continuous time ranges. For example, the second time range may precede the first time range.

For example, the filter kernel width and / or the filter kernel length are adjusted stepwise by, for example, a fixed number of frequency beams or a fixed number of time-blocks.

For example, the signal is obtained based on two-dimensional frequency and time filtering of the gain function, the filtering is performed with a test filter kernel, and the width and / or length of the test filter kernel And the width and length of the filter kernels currently used in the step.

Adjusting the filter kernel width and / or length is desirable for the following reasons.

The characteristic time-frequency patterns of respiration signals can vary over time and through certain people. For example, a person can breathe and breathe, where most of the breathing energy is concentrated at a particular time span, and for others, breathing energy may spread over a longer time span. Similarly, the frequency characteristics may also vary through different people. In addition to differences across people, there may also be differences in the time-frequency characteristics of a single person's respiration signal over time. For example, when a person has a cold, in general, the breathing process occurs through the mouth instead of the nose, and the time-frequency characteristics are likely to change.

Although the musical tones caused by spectral subtraction may be minimized or eliminated by selecting a sufficiently large kernel size, too large a kernel size may cause distortion of breath sounds, such as smearing out in frequency spectrum and / or time It is possible. Thus, adjusting the filter kernel width and / or filter kernel length may improve noise reduction and minimize distortion of breath sounds.

In another aspect of the present invention, there is provided a system for processing respiratory signals,

A noise reduction unit for calculating an output spectrum signal by performing a noise reduction operation on the spectral respiration signal, said noise reduction operation using spectral subtraction; And

- a filter unit or filter which performs two-dimensional frequency and time filtering of the gain function used in said spectral subtraction.

For example, the system may be configured to perform a method of processing respiratory signals as described above.

For example,

A spectral analyzer or a spectrum analyzer for calculating the spectral respiration signal based on a respiration signal, i.e. a respiration signal in the time domain.

For example, the breathing signal is input to a spectrum analyzing unit. For example, the output of the spectral analysis unit including the spectral respiration signal is input to a noise reduction unit.

For example, the output spectrum signal is output by the noise reduction unit. For example, the noise reduction unit includes a filter unit. For example, the filter unit is an intermediate filtering unit that performs two-dimensional frequency and time-intermediate filtering of the gain function. For example, the spectral analysis unit, the noise reduction unit and / or the filter unit are part of a spectral subtraction unit of the system for filtering the respiration signal.

For example, the system further includes a combining unit for calculating an output signal based on the conversion of the output spectrum signal to the time domain.

For example, the output of the spectrum analyzing unit is coupled to the input of the noise reduction unit. For example, the output of the noise reduction unit is coupled to the input of the synthesis unit.

For example, the system further includes a breath analysis unit that calculates at least one breathing parameter based on the output spectral signal. For example, the output of the noise reduction unit or the output of the spectral subtraction unit or the output of the synthesis unit may be coupled to the input of the breath analysis unit.

In one embodiment, the system further comprises a filter adjustment unit for adjusting the filter kernel width and / or the filter kernel length of the filter kernel used in the two-dimensional frequency and time filtering of the gain function, Is based on the ratio between the signal energy contained in the first time-frequency domain of the signal obtained based on the dimension frequency and time filtering and the noise energy contained in the second time-frequency domain. The filter adjustment unit may be configured to perform the step of adjusting the filter kernel width and / or the filter kernel length as described above.

These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

1 schematically illustrates a system for processing respiratory signals, and a data flow in a system.
2 illustrates the acoustic energy of the output signal of the comparative example and the respiration signal of the output signal of the system of Fig.
Fig. 3 schematically illustrates a setup for acquiring a respiration signal of a sleeping person; Fig.
4 schematically illustrates another embodiment of a system for processing respiratory signals;

Figure 1 illustrates, by way of example, a system for processing respiratory signals. A system and method for processing respiratory signals will be described below with reference to FIG.

In Figure 1, the signal or data flow is schematically illustrated by the arrows. A thin line arrow illustrates one or more parameters, if signaled in the time domain, or explicitly referred to. Thick contour arrows symbolize the flow of data blocks in the time domain or data blocks in the frequency domain.

Respiration signals 10 captured by a microphone located at a distance of, for example, 50 cm from the person sleeping are sampled by the sampler or sampling unit 12 and arranged in time-blocks. For example, the sampling frequency is 8 kHz. A sampling rate of 8 kHz is sufficient to maintain relevant information of respiratory sounds in the respiration signal. For example, the signal is converted into consecutive blocks of B samples. For example, B = 256.

The output of the sampling unit 12 is an output of the block concatenation and windowing unit 14 which constitutes successive time-blocks or blocks of M samples by concatenating samples of the current block with samples from at least one previous block Coupled to the input. For example, samples of the current block are concatenated with samples of the previous block for M samples, where M = 2B. For example, M = 512.

For example, M samples of a block are windowed by a Hann window, also known as a raised cosine window.

The output of unit 14 is coupled to the input of spectral analysis unit 16. [ For example, unit 16 uses a FFT operation to transform a block of M samples into the frequency domain. Thus, the unit 16 outputs the spectral respiration signal 18 in the form of continuous overlapping time-blocks of the composite spectrum with M frequency-bins. The output of the unit 16 is coupled to the input of the magnitude spectrum calculation unit 20. The unit 20 converts the complex frequency spectrum of the spectrum signal 18 into a magnitude spectrum signal 22 having M / 2 + 1 frequency bins. For example, M / 2 + 1 = 257 at a frequency resolution of 15.625 Hz.

The output of the unit 20 is coupled to the input of the noise estimation unit 24 and the input of the gain function calculation unit 26.

The noise estimation unit 24 estimates the spectral noise floor and outputs the magnitude noise spectrum signal 28 to the gain function calculation unit 26. [ Thus, the output of unit 24 is coupled to the other input of unit 26. For example, the noise is estimated by averaging over several successive time blocks of the magnitude spectrum signal 22. For example, the number of blocks of signal 22 may correspond to a time period of 10 seconds or more.

Unit 26 calculates the gain function 30 based on the magnitude spectrum signal 22 and the estimated magnitude noise spectrum signal 28. [ For example, the gain function 30 has the form of consecutive blocks of spectrum having spectral values between 0 and 1 and a block size of M / 2 + 1.

로서 계산될 수도 있고, 여기서, X는 스펙트럼 호흡 신호(18)이고, W는 추정된 잡음 스펙트럼 신호이고, n=0, ..., M/2+1이다. 여기서,

는 잡음 감소량을 제한하기 위한 소위 스펙트럼 플로어이다.

이면, 최대 잡음 감소량이 획득된다.Calculation of the gain function for spectral subtraction of noise is known, for example, as in the description of speech enhancement. Generally, the gain function is calculated such that the spectral subtraction of the noise may be performed by multiplying the spectral signal 18 with the gain function 30 in the frequency domain. For example, the spectral or gain function 30

Where X is the spectral respiration signal 18, W is the estimated noise spectral signal, and n = 0, ..., M / 2 + 1. here,

Is a so-called spectrum floor for limiting the amount of noise reduction.

, The maximum noise reduction amount is obtained.

The spectral or gain function 30 output by the unit 26 is input to a filtering unit 32, for example, a 2D intermediate filtering unit. For example, the filtering unit 32 performs a two-dimensional frequency and temporal median filtering of the gain function or spectrum 30. For example, the filtering unit 32 has an intermediate filtering kernel with a width of seven frequency bins and a length of seven time blocks. In the intermediate filtering, the extension from M / 2 + 1 frequency bins to M frequency bins is performed by mirroring the frequency spectrum. The filtering unit 32 outputs the filtered gain function 34. [

The outputs of unit 32 and unit 16 are coupled to the inputs of spectral multiplication unit 36 that multiply spectral signal 18 and filtered gain function 34 to perform a spectral subtraction of the estimated noise do. Thereby, the output spectrum signal 38 is generated.

The output of unit 36 is coupled to an input of a synthesis unit 40 that converts the output spectrum signal 38 into a time domain, for example, by performing an IFFT.

Thus, the units 20, 24, 26, 32, and 36 form a noise reduction unit that performs a noise reduction operation on the spectrum signal 18 to calculate the output spectrum signal 38.

The output of unit 40 is coupled to an input of a block overlap and adder unit 42 that calculates blocks of B samples based on the addition of overlapping portions of the blocks of M samples received from unit 40 . For example, the blocks of the B samples are contiguous blocks that do not overlap. The blocks output by the unit 42 are converted by the converter unit 44 into a sequence of successive samples forming the output signal 46. For example, the converter unit 44 performs down-sampling. Thus, the output signal 46 may have a sampling frequency that is lower than the sampling frequency of the sampling unit 12.

For example, the output of the converter unit 44 is input to the respiration signal evaluation unit 48, which calculates at least one breathing parameter 50 based on the output signal 46. The respiratory parameter may be, for example, the respiratory rate. For example, the respiration signal evaluation unit 48 calculates the respiration rate from the output signal 46 in a manner known in the art.

For example, a system for processing respiratory signals may include a processor executing a computer program configured to perform the methods described above. For example, a processor and / or computer program may be implemented as part of one of the aforementioned units, such as units 12, 14, 16, 20, 24, 26, 28, 32, 36, 40, 42, 44 and / And / or may be formed. For example, the processor and / or computer program may be part of a computer configured to perform the methods described above.

Figure 2 shows a diagram (A) depicting the acoustic energy of a signal over time. The acoustic energy is shown in any unit that refers to the signal values output by the sampling unit 12 of the system of FIG.

Figure (A) shows a real-life record of the respiration signal of a person breathing over approximately 24 seconds. The breathing signal is recorded using a microphone located approximately 50 cm from the breathing person. The sample rate was 800 Hz and values of B = 256 and M = 512 were used. As can be seen in the figure, the signal-to-noise ratio is poor, and the sounds of breathing are invisible.

As a comparative example, Figure B shows an output signal obtained from the system as shown in Figure 1, but not performing the intermediate filtering by the unit 32. [ Thus, Figure (B) shows the result of the conventional spectral subtraction of the estimated noise. It has been found that the spectral subtraction method produces musical tones in the output signal with very stochastic characteristics both in time and frequency. Some of these tones or outliers can be seen in FIG. These outliers can be easily misclassified as respiration by a respiratory signal evaluation unit or procedure.

FIG. 5C shows the result of processing the breathing signal of FIG. A with the system of FIG. 1 including a 2D intermediate filtering unit 32. The kernel size of the seven frequency bins x seven time-blocks was used to obtain the output signal 46 shown in Fig. As can be seen, the voices are efficiently removed. Thus, classification and extraction of respiratory parameters by the unit 48 can be improved.

Since the respiration increases and decreases very slowly with time and since the frequency content of the respiration also does not indicate very large peaks and valleys present in the vowels of speech, May be performed in two dimensions without excessive distortion.

FIG. 3 schematically shows a possible setup for capturing the breathing signal 10 using one or more microphones 52 disposed in the vicinity of the sleeping person 54. For example, two microphones 52 are placed in the front corner of the bed. Using two microphones at different positions allows acceptance of changes in the positions of the sleeping person 54. [ For example, the method described above may be performed on two respiration signals 10 corresponding to different microphones 52, and a respiration signal having a higher signal-to-noise ratio may be used to calculate at least one respiratory parameter . ≪ / RTI >

) 및 필터 커널 길이(

)를 결정한다. 유닛(56)의 파라미터 출력들은 유닛(32)의 파라미터 입력들에 커플링된다.4 illustrates another embodiment of a system for processing respiratory signals that is different from the system of Fig. 1 in that the noise reduction unit further includes a filter adjustment unit 56. Fig. The output of the gain function calculation unit 26 is coupled to the input of the filter adjustment unit 56. The filter adjustment unit 56 is configured to adjust the filter parameters to be used by the filtering unit 32,

) And filter kernel length (

). The parameter outputs of unit 56 are coupled to the parameter inputs of unit 32.

는 주파수에 대한 모델 차수이고

는 시간에 대한 모델 차수이다.The kernel size corresponding to the width of the plurality of frequency bins and the length of the plurality of time blocks may be interpreted as the model degree of filtering. therefore,

Is the model order for the frequency

Is the model order for time.

) 및 필터 커널 길이를 조정하기 위해, 유닛(56)은 아래의 계산 단계들을 수행한다.The filter kernel width of the filter kernels used by unit 32 (

) And the filter kernel length, the unit 56 performs the following calculation steps.

Unit 56 determines a first time range to calculate the desired signal, i. E., The signal energy of the breathing sound. For example, the detection moment t at which the signal energy of the gain function reaches a maximum is determined. For example, the first time range is set to [t-0.25, t + 0.25], where time values are provided in seconds.

For example, a second time range is determined to calculate the signal energy of the noise signal. For example, the second time range is set to [t-0.75, t-0.25].

For example, the two-dimensional frequency and time filtering of the gain function is then performed using a test filter kernel having the same kernel size currently used by unit 32 to generate the filtered signal.

For example, signal energy included in a first time range of the signal and a first or lower frequency range of, for example, 0 to 2 kHz is calculated and a second or higher frequency of 2 to 4 kHz The signal energy contained in the range is calculated, and the ratio between the two signal energy values is calculated.

If this ratio is greater than a certain threshold, it is assumed that there is a breathing sound. However, if the ratio is less than the threshold value, this indicates that the described musical tones are in the higher frequency range. Therefore, the noise reduction is insufficient.

) 및 커널 필터 길이(

)가 필터링 유닛들(32)로 출력된다.In the case of insufficient noise reduction, for example, filtering the gain function with the test filter kernel and determining the rate may be repeated for one or more test filter kernels of different kernel sizes, The size may be selected. Thereafter, the selected kernel filter width (

) And kernel filter length (

Are output to the filtering units 32. [

의 테스트 필터 커널들을 사용하여 결정될 수도 있고, 여기서,

는 현재 커널 사이즈이다. 예를 들어, 최적의 비율을 갖는 커널 차수가 새로운 커널 사이즈로서 선택될 수도 있다.For example, the ratio may be a function of kernel sizes or kernel orders

May be determined using test filter kernels of < RTI ID = 0.0 >

Is the current kernel size. For example, a kernel order with an optimal ratio may be selected as the new kernel size.

For example, the above-described steps of determining the first and second time ranges and determining the ratio for each of the time ranges may be repeated for a plurality of breaths, and the ratio may vary over the plurality of breaths May be averaged.

For example, the unit 56 may lower the model order for both time and frequency, i. E. After a certain time has elapsed, for example, every 10 minutes, the filter kernel width and length may be, for example, Value. Thus, leaks of model orders over time may be applied. This is because when only the test filter kernels having a larger filter kernel width and / or a larger filter kernel length than the currently used filter kernel in unit 32 are tested in determining the ratios by using test filter kernels of different sizes .

Adjusting the filter kernel width and / or the filter kernel length as described above allows determination of the optimal filter kernel size, human respiratory condition, and / or acoustic condition for the individual person. Thus, the calculation of breathing parameter (s) 50 may be improved.

While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description should be regarded as illustrative and not in a limiting sense. The present invention is not limited to the disclosed embodiments.

Modifications to the disclosed embodiments may be understood and effected by those skilled in the art in practicing the claimed invention from a review of the drawings, the disclosure, and the appended claims.

In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. Any reference signs in the claims should not be construed as limiting the scope.

Claims (14)

CLAIMS What is claimed is: 1. A method for processing respiratory signals,
- performing, on the spectral respiration signal (18), a noise reduction operation using spectral subtraction to compute an output spectral signal (38); And
- performing two-dimensional frequency and time filtering (32) of the gain function (30) used in said spectral subtraction of performing said noise reduction operation,
Wherein the gain function (30) is calculated based on the estimated noise and the magnitude spectrum signal calculated based on the spectral respiration signal, and wherein the two-dimensional frequency and time filtering of the gain function is at least 0.05 second time- Of the frequency-span of the breathing signals. The method according to claim 1,
Wherein the two-dimensional frequency and time filtering (32) of the gain function (30) is two-dimensional frequency and time-averaged or median filtering. 3. The method according to claim 1 or 2,
- calculating the spectral respiration signal (18) based on the respiration signal (10). 3. The method according to claim 1 or 2,
- calculating at least one breathing parameter (50) based on said output spectral signal (38). 3. The method according to claim 1 or 2,
The noise reduction operation may include:
- calculating a magnitude spectrum signal (22) based on said spectral respiration signal (18);
- calculating a gain function (30) based on the estimated noise (24) and the magnitude spectrum signal (22);
Performing two-dimensional frequency and time filtering (32) of the gain function (30); And
- calculating the output spectrum signal (38) by multiplying the spectral respiration signal (18) with the filtered gain function (34). 3. The method according to claim 1 or 2,
Further comprising adjusting (56) at least one of a filter kernel width and a filter kernel length of the filter kernel used in the two-dimensional frequency and time filtering (32) of the gain function (30)
Wherein the adjusting is based on a ratio between signal energy included in a first time-frequency domain of a signal obtained based on two-dimensional frequency and time filtering of the gain function and noise energy included in a second time- / RTI > The method of claim 1, further comprising: 21. A system for processing respiratory signals,
- a noise reduction unit (20, 24, 26, 32, 36) for calculating an output spectrum signal (38) by performing a noise reduction operation on a spectrum respiration signal (18), said noise reduction operation using spectral subtraction , The noise reduction unit (20, 24, 26, 32, 36); And
- a filter unit (32) for performing two-dimensional frequency and time filtering of the gain function (30) used in said spectral subtraction,
Wherein the gain function (30) is calculated based on the estimated noise and a magnitude spectrum signal calculated based on the spectral respiration signal, and wherein the two-dimensional frequency and time filtering of the gain function is at least 0.05 second time- Wherein the filter kernel is implemented with a filter kernel corresponding to a frequency-span of 40 Hz. 8. The method of claim 7,
- a spectral analysis unit (16) for calculating the spectral respiration signal (18) based on a respiration signal (10). 9. The method according to claim 7 or 8,
- a filter adjustment unit (56) for adjusting at least one of a filter kernel width and a filter kernel length of a filter kernel used in the two-dimensional frequency and time filtering (32) of the gain function (30)
Wherein the adjustment is based on a ratio between signal energy included in a first time-frequency domain of a signal obtained based on two-dimensional frequency and time filtering of the gain function and noise energy included in a second time- A system for processing respiratory signals. A computer-readable recording medium having recorded thereon a computer program for executing the method according to any one of claims 1 to 3 when executed on a computer. delete delete delete delete

Images